Re: [PATCH] smc: fix refcount bug in sk_psock_get (2)

[Hawkins Jiawei]

On Sat, 9 Jul 2022 at 11:06, Jakub Kicinski <kuba@xxxxxxxxxx> wrote:
> On Sat,  9 Jul 2022 10:46:59 +0800 Hawkins Jiawei wrote:
> > Reported-and-tested-by: syzbot+5f26f85569bd179c18ce@xxxxxxxxxxxxxxxxxxxxxxxxx
> > Signed-off-by: hawk <18801353760@xxxxxxx>
> > ---
> >  net/ipv4/tcp.c | 13 +++++++++++++
> >  1 file changed, 13 insertions(+)
> >
> > diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> > index 9984d23a7f3e..a1e6cab2c748 100644
> > --- a/net/ipv4/tcp.c
> > +++ b/net/ipv4/tcp.c
> > @@ -3395,10 +3395,23 @@ static int do_tcp_setsockopt(struct sock *sk, int level, int optname,
> >       }
> >       case TCP_ULP: {
> >               char name[TCP_ULP_NAME_MAX];
> > +             struct sock *smc_sock;
> >
> >               if (optlen < 1)
> >                       return -EINVAL;
> >
> > +             /* SMC sk_user_data may be treated as psock,
> > +              * which triggers a refcnt warning.
> > +              */
> > +             rcu_read_lock();
> > +             smc_sock = rcu_dereference_sk_user_data(sk);
> > +             if (level == SOL_TCP && smc_sock &&
> > +                 smc_sock->__sk_common.skc_family == AF_SMC) {
>
> This should prolly be under the socket lock?
>
> Can we add a bit to SK_USER_DATA_PTRMASK and have ULP-compatible
> users (sockmap) opt into ULP cooperation? Modifying TCP is backwards,
> layer-wise.

Thanks for your suggestion, I also agree that modifying TCP directly
is not wise.

I am sorry that I can't follow you on haveing ULP-compatible
users (sockmap) opt into ULP cooperation, yet adding a bit to
SK_USER_DATA_PTRMASK seems like a good way.

I plan to add a mask bit, and check it during sk_psock_get(),
in v2 patch

>
> > +                     rcu_read_unlock();
> > +                     return -EOPNOTSUPP;
> > +             }
> > +             rcu_read_unlock();
> > +